What do you want Jev to say?
Jev is a decision model. It can't write. Every word here is it answering a multiple-choice question about what the next word should be. How, and why. Bring your own key from typesafe.ai.
A decision model that can't generate text, made to chat one word per round trip.
Jev is a decision model. It can't write. Every word here is it answering a multiple-choice question about what the next word should be. How, and why. Bring your own key from typesafe.ai.
What Jev is for. Jev is TypeSafe's "System One" model. You send it some state and typed questions: pick one of these options, score this against a rubric, is this statement true. It answers with calibrated probabilities, in under a second, hundreds of questions per call. It is built to be the part of your software that decides things: which team gets the ticket, whether the review is spam, which of 200 products a description matches. It has no output channel for text. It cannot say anything.
What this site does with it. It asks Jev, over and over, "here is the reply so far; which of these words comes next?" Jev caps a question at 255 options, so the vocabulary is split across several questions, the best guesses from each are pooled, and a second question shows Jev the actual sentences ("...bring the water to a boil" vs "...bring the water to a go") and asks which reads best. Then it appends the winner and asks again. A speculative round guesses the following word in the same call so most words cost one round trip. Punctuation, ending, capital letters and a/an are all either more questions or rules, because Jev has opinions about none of them.
How bad an idea this is. A normal language model produces a word in a single forward pass. This produces a word with roughly sixty multiple-choice questions, about 1.3 seconds, and a fraction of a cent. It can't spell, so any word that isn't in its 4,000-word list or in your message is simply unavailable to it. It can't quite tell when it's finished. It has never seen the sentence it is writing as anything but a list of options. Also it wrote a usable egg recipe on the first try, which is more than the setup deserves.
Code, benchmarks and everything that didn't work are on GitHub.
Sent straight to api.typesafe.ai for your own requests and kept in this browser's local storage. Nothing is logged server-side. A 30-word reply costs about a cent.